Noise (I assume) and what to do with it

[DCK]

Hi all,

I've been using PhotoScan to make 3D models of skulls. The results are spectacular in some ways. However, I am having two issues:

1) Where there are holes in the skull, the model output always has lots of junk. I assume this is "noise." If not, please tell me.

2) When I analyze the model vs. the output of a laser surface scanner, the PhotoScan surface area estimates tend to be larger. I've raised this in a separate post ("Model Surface Area")

So, back to the noise (Issue 1). I have attached a jpeg with an example of what I'm talking about. There are three images of the model, as well as one masked photograph. The area I am most concerned with is called the "temporal fossa." I have circled it in blue on all the images. You can see on the photograph that I masked this region precisely. I do so in all the photos. It does not make a difference.

Also, in the second model image, you will see that I have circled that nasal aperture (hole) in red. Unlike the temporal fossa, it does not pass all the way through (or else you'd have a hole in the back of your head). But it has the same "noise" problem.

Is there a way to eliminate this issue. I know I can clean the model after I've built geometry/texture. However (1) this is time consuming, and (2) since "noise" triangles inevitably attach to triangles that are part of the actual anatomy, the editing introduces inaccuracies and undesirable hole-filling.

Thanks much.

DCK

[Wishgranter]

Can post screen from pscan to see camera positions ? the side where is the messy stuff.....

[DCK]

Sure Wishgranter, but I would be very surprised if this is about photo coverage. I take about 100 (90-120) photos per specimen and am pretty diligent about getting coverage of hard to spot areas. Also, the nasal aperature and eye orbits face forward, and they too have noise (see red circle on first image).

[Wishgranter]

send me PSZ file without mesh, just aligned cameras, its highly possible that you have some photos near each other, say just 1-3cm separation......

send it to muzeumhb@gmail.com 

[DCK]

On it's way. I've been more concerned with lack of coverage than overlap, so maybe you're right.

Thanks.

David

[jedfrechette]

Have you tried exporting a dense point cloud and inspecting it in something like CloudCompare?

For complicated objects I don't have laser scans of I've been moving towards exporting the point cloud from PhotoScan and putting it into my normal laser scanning pipeline. Edit point cloud in PolyWorks, generate mesh with PoissonRecon and PolyWorks, UV unwrap, then reimport into PhotoScan to lay down the base texture.

[DCK]

Interesting. I've wondered if differences in mesh generating algorithms (?) between the laser scanning software and PhotoScan could generate differences in surface area (see my other post "Model Surface Area"), but I haven't tried it.

I will see what I can do. Unfortunately, I'm traveling so it will be a couple of weeks before I'm able to give this a try.

[jedfrechette]

I don't know that it's documented anywhere but from looking at the results my guess is that PhotoScan uses some sort of a Poisson Surface Reconstruction (PSR) internally when building arbitrary smooth meshes, so I'm not sure it's a case of different algorithms. I just prefer to go straight to the source (http://www.cs.jhu.edu/~misha/Code/PoissonRecon) in order to get more control over the results. The Kazhdan and Bolitho version also gets cool new features before they show up in any of the third party tools that are based on their work. For many data sets, PSR is also much better than any of the algorithms, proprietary or otherwise, that I've found in commercial laser scanning packages.

Regardless, looking at the raw point cloud should provide some insight into how severe the noise is and perhaps where it is coming from.

[Maharg]

Do the initial alignment, export the cameras xml and force your camera's calibration that was calculated.  Run your photo sequence through neat video (the most amazing degrainer available) and render it with the same naming/numbering convention as imported.  Swap the folders so Photoscan opens the degrain sequence instead.  Import the Cameras xml and resolve.  I've gotten double the points with this process on numerous occasions.

[lunar]

i have the same problem, and i doing like this:

strange is when i reconstruct geometry using "sharp" i get clean model i the previosly noise places (but with other noise and holes from "sharp" processing)


after photo aligment -> build geometry -> low, smooth
then after first model is ready i inspect it, and looking for noise places (i always find some of them - noisy craters) , then i turn in to depth map view and i search for noisy Depth Maps of this area and mask or block them.
After that i generate model again - now is clean.
Some times i have to repeat that steps to get nice & clean model

the problem seems to heve something with depth map generation.


I also using lot of photos(>200 - close up`s for the best texture) to reconstruct full 360 a skull size objects